Metaphor Suggestions based on a Semantic Metaphor Repository

Metaphor Suggestions based on a Semantic Metaphor Repository Gerard de Melo Rutgers University – New Brunswick Department of Computer Science Piscataway, NJ, USA gdm@demelo.org Abstract Metaphors are not only remarkably pervasive in both informal and formal text, but reflect fundamental properties of human cognition. This paper presents an algorithmic model that suggests metaphoric means of referring to concepts. For example, given a word such as government, the method may propose expressions such as father, nanny, corresponding to common ways of thinking about the government. To achieve this, the model draws on MetaNet, a manually created repository of conceptual metaphor, in conjunction with lexical resources like FrameNet and WordNet and automated interlinking techniques. These resources are connected and their overall graph structure allows us to propose potential metaphoric means of referring to the given input words, possibly constrained with additional metaphoric seed words, which may be provided as supplementary inputs. The experiments show that this algorithm greatly expands the potential of the original repository for this task by enabling new connections to be drawn. Keywords: Metaphor, lexical resources, graph structure 1. Introduction Whenever one says that issues become clear, stock markets go up, or time is spent, language is arguably being used in a non-literal, metaphorical manner, at least with respect to the original senses of the words. Corpus studies have found that metaphorical phenomena are very pervasive even in formal language (Steen et al., 2010; Shutova and Teufel, 2010). Not only is such metaphorical use of language one of the primary means for creative linguistic expression. It has been widely stipulated that our reliance on metaphor is a natural consequence of the way our brains reflect on and reason about the world. This paper presents a model that can be used to suggest both well-entrenched and novel metaphoric means of referring to a given input word or set of related input words. For ex- ample, given a word such as government, the method may propose expressions such as father, nanny, corresponding to ways of thinking about the government. While metaphor has been studied extensively in linguis- tics, cognitive science, as well as NLP, the task of auto- matically suggesting metaphors has received only little at- tention. Young (1987) relied on simple relational database queries to find related words. Abe et al. (2006) proposed a method that takes a noun and a set of adjectives as input (e.g., person and young, innocent, fine) and use corpus topic models to find other nouns with these properties. Veale and Hao (2007) extend this idea to Web-scale knowledge by using the Google search engine to find relevant adjec- tives describing a noun. Terai and Nakagawa (2010) present an alternative model for this, based on semantic similarity, which considers both adjectives and verbs as relevant noun properties. Approaches of this sort excel at finding novel poetic metaphors, e.g. hope is like a lightbulb, as discussed by Terai and Nakagawa (2010). The system discussed in this paper, in contrast, is biased towards finding novel varia- tions …

Content truncated to 500 words